GDPval

Sharks, Lee

SPXI instrument · method 0.9 of 87, level 0 (Measurement): Capture Registry (longitudinal observation). An instrument of SPXI, the Semantic Packet for eXchange & Indexing; the whole practice is itemized at spxi.dev/methods.

Capture Registry › capture gdpval-aio-20261029

One record of the canonical Capture Registry (EA-WG-CAPTURES-01), cited at https://www.alexanarch.org/captures/gdpval-aio-20261029/. the canonical Capture Registry (version 12.396) · the address page · this card in the gallery · this record as data · table of contents.

AXN:0AED.CAPTURE — this address. Each observation below is one AXN, sealed with its analysis and measures; the identifier names the reading together with the response it reads.
AXN:0AED.CAPTURE.🕐🕕▶️🗡️⌛🪧 — Google AI Overview, 2026-10-29 · sealed observation (SHA-256 6267d35a604de8b3…)

Machine Reception2026-10-29

GDPval

Originator: OpenAI: the paper's 19 authors (Tejal Patwardhan first; nine marked equal contribution) give OpenAI as their affiliation, with correspondence to an openai.com address; OpenAI's own post at this address answered 403 and is unread · outside the archive (out-universe) · dataset · SPXI treatment: partial

BROAD MATCH

no
image

GDPVAL, BY NAME, ON GOOGLE'S AI OVERVIEW, SIGNED OUT, 2026-10-29: 'an AI benchmark created by OpenAI to measure how well artificial intelligence models perform on real-world, economically valuable knowledge-work tasks rather than academic or math tests.' Scope (1,320 tasks, 44 occupations, the top 9 U.S. GDP sectors), task design by professionals averaging 14 years' experience, blind head-to-head expert grading, the 220-task gold subset on Hugging Face and arXiv, and Artificial Analysis's GDPval-AA leaderboard. Seven cards: two OpenAI pages, arXiv, Hugging Face, Artificial Analysis, Inspect Evals, a video.

Full record — 6,180 characters, 7 sources

Capture record

captured: 2026-10-29
surface: Google AI Overview
auth state: signed out, incognito
evidence class: paste
PER (dimensional, rubric v1.1): PER 0.11 · M 0.11 · C — · D 0.11
PER (historical scalar): 0.25
PER units retained: inst, id, src
citations read: 7
observation id: OBS-ef64b7d44199
address id: ADDR-d4b480b48aca

Reading

The innocent reading first. The Overview gives the benchmark's maker, scope, construction and grading, and where to find its data, from OpenAI's pages and the paper.

Findings

THE BENCHMARK EXACT. Every figure the composition gives is the record's: 1,320 tasks, 44 occupations, 9 sectors, the 220-task gold subset, experts averaging 14 years. Checked 2026-10-29 against the Labor Vector census record of the same day and the sources it saved.
'BLIND'. The record's stated limits include that grading may not have been fully blind; the composition's 'blind head-to-head comparisons' gives the design without the caveat.
THE FINDINGS ABSENT. The paper's own: models 'beginning to approach parity' (47.6% of Claude Opus 4.1's deliverables graded as good as or better than the expert's); failures to follow instructions as the commonest reason experts preferred the human work; and, once review and redo time is counted, 'the payoff from using a model shrinks' (for GPT-5, 90 times as a naive speed figure, 1.12 if tried once).
THE LIMITS ABSENT. Self-contained computer work only; precisely specified, one-shot tasks; the automated grader agreeing less with experts on capable OpenAI models. The record's negative tags keep the benchmark apart from a measure of use or employment; the composition claims neither.
THE EXPERTS' TERMS. Who wrote and graded the tasks is given ('industry professionals'); how they were recruited and paid, which the paper gives only as 'well compensated', is not composed.

Analysis analyst prose, not machine text

A capability benchmark is composed at its specification and stops before its results. The record's labour question, the experts whose work product the tasks are built from and the review time charged against the model, is where the paper meets user labour drag; the composition carries the first as design and leaves out the second.

Machine text, verbatim

CAPTURE-TIME VERBATIM RECORD — paste of the Overview expanded, with its source cards · complete as supplied · READ IN FULL 2026-10-29

[Google AI Overview at the query, signed out and incognito (operator defaults of 2026-10-01 and 2026-10-14), 2026-10-29. The paste is the Overview expanded (the 'AI Mode Conversation' view, with the source cards that follow the answer); the 'You said' line carries the query twice, run together, and is given once here. Inline links are Google redirect URLs, kept as pasted.] [QUERENT] GDPval [SYSTEM] GDPval is an AI benchmark created by [OpenAI](https://www.google.com/goto?url=CAESWAHrOzAVhPUzkeZH77Ag5jLMmksSXvMZIZn3Sp_gF5kA_w6WuRy05O9W5aaR91vp2mcMDj0LP6Z6Dh3R6CNeLyRGrDRztzxt5b0SrhXSwBWUrWI8BUM3oPg) to measure how well artificial intelligence models perform on real-world, economically valuable knowledge-work tasks rather than academic or math tests. [[1], [2], [3]] Key Details * Scope: Covers 1,320 specialized tasks across 44 occupations derived from the top 9 U.S. GDP-contributing sectors. [[1], [2]] * Task Design: Built by industry professionals with an average of 14 years of experience using actual work products like legal briefs, spreadsheets, and engineering blueprints. [[1], [2]] * Evaluation: Uses blind head-to-head comparisons by human experts to grade model outputs against human work. [[1], [2]] * Open Source Set: Includes a gold open-sourced subset of 220 tasks available via [Hugging Face Datasets](https://www.google.com/goto?url=CAESZQHrOzAVm5FY12OrrM0U6PSxauyQdUwzN8-8rk1Z-9EU8DSkpEJyiaghIjM8OUDTTTwYY9F3FOr1suUN7D-RQdFC44Y8DZqzlNkNd4_3MEsmBpgPdNQl0PKHVzYpyxlHs1qyhSgQ) and detailed in the [ArXiv Research Paper](https://www.google.com/goto?url=CAESWAHrOzAV5e1D92fjl4arLGgbWubPM_woQLTk88wKqtsjeycDpo4ugmHlLKBY2Jmgp8L30J-U1cjOBDae-FiGJbaIH7bL25Gw_j2LadpPeqBjk_YuRdt18qQ). [[1], [2]] * Leaderboards: Evaluated by frameworks like [Artificial Analysis GDPval-AA v2.1](https://www.google.com/goto?url=CAESawHrOzAVgAjm8sR1LVSiS9i7GzX2qM_VLeqRumJBavESmc8WxlV73ea4yA7gFz_me9eppW0Og-Ml_WxAMVWD3bWEpsY0O2plbNqtIGV7xMNVdQ7190-HFX_X5hjx_vrKAGEnxHNh9xe-HRA4). [1] Would you like to know more about top-performing models on GDPval, or how to access the open-source dataset? OpenAI

Sources (7) as cited, and as the copy produced them

OpenAI third_party
Measuring the performance of our models on real-world tasks | OpenAI
GDPval: Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries (threshol...
as pastedMeasuring the performance of our models on real-world tasks | OpenAI GDPval: Evaluation measuring model performance on 1,320 specialized tasks across 44 knowledge work occupations from 9 U.S. GDP-contributing industries (threshol... Artificial Analysis
Artificial Analysis third_party
GDPval-AA v2.1 Benchmark - Artificial Analysis
GDPval-AA v2.1 Benchmark LegacyAgenticKnowledge WorkWritingReasoningInstruction FollowingTool UseMultimodalBusinessFinanceLegalMedicalHumanities GDPval-AA v2. 1...
as pastedGDPval-AA v2.1 Benchmark - Artificial Analysis GDPval-AA v2.1 Benchmark LegacyAgenticKnowledge WorkWritingReasoningInstruction FollowingTool UseMultimodalBusinessFinanceLegalMedicalHumanities GDPval-AA v2. 1... arXiv
arXiv third_party
[2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
[2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks Scheduled maintenance: Today - Please check here for updates:Stat...
as pasted[2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks [2510.04374] GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks Scheduled maintenance: Today - Please check here for updates:Stat... OpenAI
OpenAI third_party
gdpval: evaluating ai model performance - OpenAI
Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks. GDPval covers the top 9...
as pastedgdpval: evaluating ai model performance - OpenAI Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks. GDPval covers the top 9... Hugging Face
Hugging Face third_party
openai/gdpval · Datasets at Hugging Face
Dataset: openai/gdpval on Hugging Face (Parquet format, <1K size, 220 rows, 2.29 GB total file size). Modalities: Text. Libraries: Datasets, pandas, Polars, xet...
as pastedopenai/gdpval · Datasets at Hugging Face Dataset: openai/gdpval on Hugging Face (Parquet format, <1K size, 220 rows, 2.29 GB total file size). Modalities: Text. Libraries: Datasets, pandas, Polars, xet... YouTube·Vivek Haldar 11:23
YouTube·Vivek Haldar third_party
GDPval: An AI benchmark for real-world tasks
video, 11:23
as pastedGDPval: An AI benchmark for real-world tasks GitHub Pages documentation
GitHub Pages documentation third_party
GDPval – Inspect Evals - GitHub Pages
GDPval measures model performance on economically valuable, real-world tasks across 44 occupations. The first version of this evaluation spans 44 occupations se...
as pastedGDPval – Inspect Evals - GitHub Pages GDPval measures model performance on economically valuable, real-world tasks across 44 occupations. The first version of this evaluation spans 44 occupations se...

↻ Re-run permalink